Unsupervised Change Point Detection Using a Weight Graph Method for Process Monitoring

Abstract

Because industrial processes are complicated and time-varying in general, unsupervised, and nonparametric process monitoring methods are necessary. Recently, a graph-based change point detection method with a developed scan statistic, which is unsupervised and nonparametric, has been introduced. Industrial processes are primarily continuous with considerable important information contained in the time relations of adjacent observations. This important information should be used for process monitoring, which could improve the power of change point detection. In this paper, the scan statistic based on the graph method is adopted for process monitoring and the time intervals between observations are attempted for calculating the weights of the edges. There are two steps for detecting the change point. (1) Construct the connection graph: a minimum spanning tree is used for constructing the connecting graph, and the weights of the edges are calculated based on the time intervals and Euclidean distances between the observations. (2) Calculate the scan statistic: the number of edges connecting the observations derived from two parts (before the change point and after the change point) is counted as a statistic for detecting the change point. In this paper, the Tennessee Eastman (TE) process and a blast furnace process are used to illustrate the power of the weight graph method for process monitoring.

Graph-Based Method for Fault Detection in the Iron-Making Process

“…The poor quality of the raw materials, abnormal airflow distribution and irregular furnace wall are the main reasons for the slip. Slip can be solved by appropriately reducing the wind volume [9], [15].…”

“…Escobar et al proposed a combined generative topographic mapping and graph theory approach for unsupervised nonlinear data visualization and fault identification [14]. An et al adopted an unsupervised graph method for fault detection, and the time interval was used to improve the power [15]. A support vector clustering-based probabilistic approach was developed for unsupervised chemical process monitoring and fault classification by Yu [16].…”

“…An et al used the time interval to improve the process monitoring power based on the graph method. However, the relations between variables are not considered, which may be the reason why the method is ineffective for some faults [15]. The graph-based change point detection method uses the Euclidean minimum spanning tree to build the graph.…”